Using AI in Medicine: the Slow AI Guidelines

A strange pain, an insect bite, a question about food… what do you do? Ask an AI. For many people, it’s now an automatic reflex.

Generative artificial intelligence tools such as ChatGPT, Claude, and Gemini are increasingly being used by the public as sources of health information.

This is why Slow AI, a project by Slow Medicine ETS, was created: to provide guidance that helps both the public and healthcare professionals use AI appropriately when seeking or providing health information, while discouraging its inappropriate use.

The project was carried out by a multidisciplinary group of experts from Zadig, the Italian Society for Artificial Intelligence in Medicine (SIIAM), the Italian Study Group for Diabetes Education (GISED), the Mario Negri Institute for Pharmacological Research, the University of Foggia, the Turin Medical Association (OMCeO), and Slow Medicine.

The project consisted of three complementary studies. These assessed the ability of generative AI systems to provide information consistent with clinical guidelines, to answer health-related questions asked by members of the public, and to explore how AI is currently being used for medical topics.

The findings generated a substantial body of evidence that made it possible to reflect on both the risks and the opportunities of AI in healthcare, and to identify what a “measured, respectful and fair” use of artificial intelligence, echoing Slow Medicine’s guiding principles, should look like when responding to citizens’ health questions.

This work ultimately led to the development of the Slow AI Decalogue: ten practical recommendations to promote the informed and responsible use of generative AI.

Slow AI Decalogue

1. When asked health-related questions, generative AI systems often produce highly convincing answers. Their responses appear authoritative, reliable, and appropriate, but this appearance of accuracy can be misleading.
2. Generative AI systems do not rely on clinical reasoning or medical understanding. Instead, they generate responses by statistically predicting the most likely sequence of words. They do not understand the meaning of what they produce, have no direct access to scientific evidence, and cannot assess what is clinically appropriate for an individual patient.
3. Compared with physicians’ answers, responses generated by AI generally provide information that is understandable, applicable to most patients, and up to date. However, they are not free from errors and may sometimes convey misleading messages about diagnosis or treatment.
4. When compared with evidence-based clinical guidelines, AI-generated answers are generally of good quality and reasonably comprehensive. Nevertheless, they are sometimes inaccurate, incomplete, or not fully aligned with current recommendations: particularly they can be outdated. For this reason, it is essential that healthcare professionals actively contribute to the development, validation, and oversight of these systems to ensure that their outputs remain consistent with the best available scientific evidence.
5. Generative AI systems may also produce so-called “hallucinations”: responses that sound plausible or even convincing but are not supported by scientific evidence.
6. AI-generated text is written in excellent Italian, making it difficult for readers to distinguish between content produced by AI and content written by humans.
7. Providing personal health information to a generative AI system in the hope of obtaining a diagnosis carries significant risks. Besides the possibility of receiving an incorrect diagnosis, which may either cause unnecessary anxiety or provide false reassurance, delaying appropriate medical care, it also raises important concerns about the privacy and security of sensitive health data.
8. Asking AI to recommend the best treatment for a medical condition may result in advice that appears technically flawless, but is clinically inappropriate. AI cannot take into account the full complexity of an individual’s medical history, current health status neither physical examination findings, thereby offering standard “advice”, which might be suitable in an ideal situation, but which is not appropriate to the specific case.
9. Asking AI to interpret medical test results is equally problematic. Laboratory values and diagnostic imaging can only be interpreted within the context of the patient’s complete clinical picture. Moreover, if the information provided to the AI is incomplete or of poor quality, for example, an unclear CT scan image, the resulting interpretation may be entirely incorrect.
10. Taken together, these findings call for caution and healthy scepticism. Since the basis on which generative AI produces its answers remains largely opaque, its responses should not be relied upon for personal medical decision-making. AI can be useful when used to ask general, non-personal questions, for example, to learn about a particular disease (a measured use), while avoiding the sharing of personal health information (a respectful approach that protects privacy and ethical principles) and users should always critically evaluate AI-generated information (a fair use); that means that it is necessary to consult healthcare professionals to interpret AI responses and confirm any medical information before acting on it.

Further information is available on the websites of Zadig and Slow Medicine.